cancer-detection-classifier

About

Convolutional-recurrent neural network for detecting cancer from cell-free DNA whole genome sequencing data.

Inputs and Outputs

1. Input is a 1kb resolution copy number (CN) profile and windowed protection score (WPS) profile for a 1Mb genomic region.
   • shape (?, 1, 1000, 2)
2. Labels are binary indicators of the classes [1,0] for [cancer, no cancer].
   • shape (?, 1)
3. Output is the sigmoid probability for the cancer class.
   • shape (?, 1)

Model Architecture

1. Convolutional layer with
   • 320 kernels
   • kernel length: 26
   • step size: 1
   • ReLU activation
   • padding: same
2. Max-pooling layer with
   • pooling size: 13
   • pooling stride: 13
   • padding: same
3. Dropout Layer with
   • rate: 0.2
4. Bi-directional long short term memory layer
   • 320 forward neurons
   • 320 backward neurons
5. Dropout Layer with
   • rate: 0.5
6. Fully connected layer with
   • 925 neurons
   • ReLU activation
7. Fully connected layer with
   • 1 neuron (output)
   • sigmoid activation

Objective Function: binary cross entropy

License

This project is licensed under the terms of the MIT license.